WEBVTT

NOTE Sentence-level transcript of https://www.youtube.com/watch?v=9JFGohx4E7U

NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/9JFGohx4E7U.html. A cue ends where the next begins, or 2 s after its last word.

s1
00:00:12.400 --> 00:00:19.080
Without further ado, we're going to bring on Zishan for for his keynote if we're all set.

s2
00:00:20.040 --> 00:00:22.040
Hey, how are you?

s3
00:00:22.040 --> 00:00:23.440
Great I'm here.

s4
00:00:23.440 --> 00:00:26.680
Yes, to the magic of the internet.

s5
00:00:26.680 --> 00:00:27.960
It's really good to see you again.

s6
00:00:27.960 --> 00:00:35.280
Zishan has you've been speaking with AI Singapore and obviously GLM is the talk of the town right now.

s7
00:00:35.280 --> 00:00:43.760
There's been an amazing booth downstairs in the expo and I think a lot of people are just very excited to hear from you on what all is going on with ZIAI.

s8
00:00:43.760 --> 00:00:46.800
So if you want to take it away.

s9
00:00:47.360 --> 00:00:55.760
Yes, sorry again for not showing up in person, but I have a team whole team coming to the town and we have a booth.

s10
00:00:55.760 --> 00:01:05.040
So you have any questions, feel free to reach out to me on X link and also you can look for my team on our booth.

s11
00:01:05.040 --> 00:01:11.400
And since I cannot see my slides, so I will need Swee to help me flip through all the slides for me.

s12
00:01:11.400 --> 00:01:13.400
Yeah.

s13
00:01:14.120 --> 00:01:14.440
You're good.

s14
00:01:14.440 --> 00:01:17.960
Here Swee, like what which slides we are in right now?

s15
00:01:17.960 --> 00:01:20.640
Yeah, so we're we're on the opening slide.

s16
00:01:20.640 --> 00:01:23.840
We're talking about intelligence and ZIAI.

s17
00:01:25.120 --> 00:01:35.600
Okay, so yeah, so you can you can see that it's the first time for us to introduce GLM 4.2 5.2 to the world and also we are going to share something

s18
00:01:35.600 --> 00:01:43.880
about Z.AI and GLM because maybe people will think Z.AI and GLM they're irrelevant, right?

s19
00:01:43.880 --> 00:01:49.880
You are your company is not G.AI or your model's called Z1 or Z2.

s20
00:01:49.880 --> 00:01:58.840
And you can find my X and our company's X account here, so you can just search for my name Zishan Lee and the Z A I org.

s21
00:01:58.840 --> 00:02:01.440
And you can follow them for the follow-ups.

s22
00:02:01.440 --> 00:02:05.479
So, yeah, we can go to the second slide.

s23
00:02:05.880 --> 00:02:12.000
Yeah, actually I cannot see the slide, so I'll try to

s24
00:02:12.000 --> 00:02:12.600
So, we we

s25
00:02:12.600 --> 00:02:15.480
Yeah, try to make sure that it is correct.

s26
00:02:15.480 --> 00:02:19.720
So, the company actually is called Zhipu.

s27
00:02:19.720 --> 00:02:23.360
Maybe some people have heard of it.

s28
00:02:23.360 --> 00:02:27.680
And the all the models, it's it's called GLM.

s29
00:02:27.680 --> 00:02:32.040
Actually, it's not a a brand name, it's a generic term.

s30
00:02:32.040 --> 00:02:38.040
So, GLM actually represent general language model pre-training with auto-regressive blank filling.

s31
00:02:38.040 --> 00:02:42.360
And that paper was published back in 2021.

s32
00:02:42.360 --> 00:02:53.320
So, actually we were the one of the first labs to do explorations on large language models at the same time with OpenAI, and Anthropic, and DeepMind.

s33
00:02:53.320 --> 00:03:00.360
And even today, we we no longer use GLM as the architecture, we still use the name GLM as our brand name.

s34
00:03:00.360 --> 00:03:11.519
So, we use GLM 5.1, 5.2, and it become like one of our like most prod uh proddest product and model.

s35
00:03:11.519 --> 00:03:16.920
And the second thing that we look for is intelligence upper bound.

s36
00:03:16.920 --> 00:03:23.720
So, in terms of intelligence, we we may feel that it's represent IQ, something something like that.

s37
00:03:23.720 --> 00:03:32.680
And when Deep Seek launched one 01 launch, people are talking about the model's capability to solve math problems, physics problems.

s38
00:03:32.680 --> 00:03:40.920
But what actually intelligence mean is not just like IQ or AIME or other other physics problem.

s39
00:03:40.920 --> 00:03:49.840
So, from GLM 4.5 to GLM 4.7, we are exploring like several things, like reasoning, coding, and agentic capabilities.

s40
00:03:49.840 --> 00:03:55.360
So, as you can see from the slides, So, we add like Yeah, the last slide.

s41
00:03:55.360 --> 00:04:00.440
Yeah, we were actually I haven't like finished that slide.

s42
00:04:00.440 --> 00:04:03.600
Yeah, so Okay.

s43
00:04:03.600 --> 00:04:08.840
Yeah, need to go back to the two two slides.

s44
00:04:08.840 --> 00:04:14.000
I don't have back I don't have a back button.

s45
00:04:15.840 --> 00:04:17.000
Thank you.

s46
00:04:17.000 --> 00:04:19.239
I would you I don't understand clickers that don't have back buttons.

s47
00:04:19.239 --> 00:04:23.760
Like why Okay, I you know, anyway, go ahead.

s48
00:04:24.919 --> 00:04:25.800
Yeah, never mind.

s49
00:04:25.800 --> 00:04:38.160
Yeah, because people want to see the GLM 5.2. They they don't want to see you like GLM 4.1 or 5. But like GLM 4.2 actually specialize in coding and genetic task as you can see from the graph because

s50
00:04:38.160 --> 00:04:45.800
there are a lot of rumors whether your model is close to mythos, fable, but actually I want to share these slides to all of you.

s51
00:04:45.800 --> 00:04:58.840
So, you can see it's somewhere between Opus 4.7 and 4.8 and we use the hardest problems like Deep Sweep, Terminal Bench 211, which was mentioned by the Open AI team

s52
00:04:58.840 --> 00:05:01.880
uh several minutes ago.

s53
00:05:01.880 --> 00:05:16.120
And all the like long horizon task and benchmark shows that the capabilities are on par with at least Opus 4.7 and it it it shows a significant improvements over

s54
00:05:16.120 --> 00:05:23.560
1.1. Also, for GLM 1.2, we add a thinking uh level called high.

s55
00:05:23.560 --> 00:05:34.320
So, because we also noticed as we move to the harder task it may consume more tokens and also we we care a lot about the token efficiency.

s56
00:05:34.320 --> 00:05:38.880
So, it's the first time we add the high level for thinking budget.

s57
00:05:38.880 --> 00:05:45.360
But even without thinking, the non-thinking model is better than the 5.1 thinking model.

s58
00:05:45.360 --> 00:05:49.360
So, I think it's a huge improvement for the open weight model.

s59
00:05:49.360 --> 00:05:56.360
That's what really impressed the world and why people are talking about GLM 5.2 lately.

s60
00:05:56.360 --> 00:05:58.960
Okay, the next slide.

s61
00:05:59.720 --> 00:06:09.880
And one one thing that I want to mention is that GLM is more more than coding model because people use it inside clock code code X, open code.

s62
00:06:09.880 --> 00:06:13.480
But actually, we have trained a lot of things outside coding.

s63
00:06:13.480 --> 00:06:20.440
For example, we improve a lot in GDP valve and also math problems.

s64
00:06:20.440 --> 00:06:23.200
We also care about math problems, frankly speaking.

s65
00:06:23.200 --> 00:06:28.480
And also, we train a lot of thing uh related to role play, general chat.

s66
00:06:28.480 --> 00:06:32.080
We want to improve every aspect of the model.

s67
00:06:32.080 --> 00:06:42.200
So, you can see from the artificial analysis intelligence index actually leads the other open way model a lot and close to the frontier model.

s68
00:06:42.200 --> 00:06:56.480
So, I want you if you are you haven't experienced GLM yet, you can use GLM to do general chat, use it to process your daily workflow, not just for coding, but you can explore the model

s69
00:06:56.480 --> 00:06:59.120
like beyond the coding scope.

s70
00:06:59.120 --> 00:07:01.400
Next slide.

s71
00:07:01.880 --> 00:07:06.560
And GLM 4.2 is a open weight model.

s72
00:07:06.560 --> 00:07:09.360
So, people always ask me, "Why do you open weight?

s73
00:07:09.360 --> 00:07:17.040
So, do you care about your business or like do you care about losing market to some infra suppliers?"

s74
00:07:17.040 --> 00:07:24.480
But actually, we open the weights for several things because there are users' needs and there are our own needs.

s75
00:07:24.480 --> 00:07:28.480
If we can meet their needs, I think it's okay.

s76
00:07:28.480 --> 00:07:31.680
It's definitely okay for us to to open the model.

s77
00:07:31.680 --> 00:07:40.360
For example, if our users want security and control and we want to build trust, we can open weight model.

s78
00:07:40.360 --> 00:07:51.680
For for some cases, if enterprise or government especially in the Western world, want to use the model, we open way, we upload to the hugging face so that they can

s79
00:07:51.680 --> 00:07:53.960
use the model on premise.

s80
00:07:53.960 --> 00:08:03.960
I think it's very beneficial for the whole ecosystem it to explore the model, especially when the capabilities is close to the frontier model.

s81
00:08:03.960 --> 00:08:07.760
And second, if people want diversity.

s82
00:08:07.760 --> 00:08:15.880
So, in terms of diversity, I mean the capabilities in legal, finance, security, they can fine-tune the model.

s83
00:08:15.880 --> 00:08:19.000
So, we need to open the model to let them fine-tune.

s84
00:08:19.000 --> 00:08:29.600
For example, Harvey is fine-tuning GLM 4.1, and maybe they're thinking about fine-tuning GLM 5.2 afterwards.

s85
00:08:29.600 --> 00:08:33.840
And I've I have like talked to a lot of other companies.

s86
00:08:33.840 --> 00:08:45.720
They're also thinking about fine-tuning GLM as their like next step or next strategy to differentiate themselves from other application company.

s87
00:08:45.720 --> 00:08:57.480
And third, if our customer or an individual want to co-design and predict the future, sometimes they need to see the architecture of the model.

s88
00:08:57.480 --> 00:09:01.760
They need to see the recipe of how you train the model.

s89
00:09:01.760 --> 00:09:05.160
So, we want to make the norm.

s90
00:09:05.160 --> 00:09:07.120
We want to make my bets.

s91
00:09:07.120 --> 00:09:13.520
So, we want to co-shape the future with our customers with the open source uh community.

s92
00:09:13.520 --> 00:09:20.240
So, I think our needs and our ecosystem really like fits into each other.

s93
00:09:20.240 --> 00:09:35.080
And GLM 5.2 couldn't succeed without you, all the open source community players like Onflos and Media and and some like individuals, super developers, they are part of it.

s94
00:09:35.080 --> 00:09:40.120
And thanks a lot to application builders like Peter.

s95
00:09:40.120 --> 00:09:40.280
Right?

s96
00:09:40.280 --> 00:09:49.840
Because open source doesn't just include open source model, but also open source softwares and open source other sorts of support.

s97
00:09:49.840 --> 00:09:59.040
So, all your support and what are you you're doing right now pushes to to make better models.

s98
00:09:59.040 --> 00:10:02.760
I think uh you you are the true hero.

s99
00:10:02.760 --> 00:10:08.720
And the last, I want to share a a great resource for you to go through.

s100
00:10:08.720 --> 00:10:11.480
Geo 5.2 is our tech blog.

s101
00:10:11.480 --> 00:10:24.520
So, actually in that tech blog, we share several things like our hugging face repo, how how you can try the Geo and 5.2. You can try the inside the chatbot

s102
00:10:24.520 --> 00:10:27.120
agent, and also you can call the API.

s103
00:10:27.120 --> 00:10:36.080
And also we have a coding plan like the Codex or cloud core subscription for you to use your tokens as individuals.

s104
00:10:36.080 --> 00:10:43.560
And also we share something about our training pipeline training recipe, which you can understand why it's a good model.

s105
00:10:43.560 --> 00:10:51.720
So, there are a lot of things behind the behind the model, not just um a a model that had great data.

s106
00:10:51.720 --> 00:10:57.560
We also have fantastic technologies behind the model, so you can export the model yourself.

s107
00:10:57.560 --> 00:11:02.200
You can see what difficulties we have gone through.

s108
00:11:02.200 --> 00:11:07.280
And the last slide, actually it's kind of a one more thing.

s109
00:11:07.280 --> 00:11:13.880
So, it's the first time we share the code to the whole community.

s110
00:11:13.880 --> 00:11:14.240
Woo!

s111
00:11:14.240 --> 00:11:19.040
So, that one more thing is we actually have our own harness.

s112
00:11:19.040 --> 00:11:26.080
The Z code, actually it's built for Geo 5.2, but also support all frontier models.

s113
00:11:26.080 --> 00:11:27.920
You can bring your own key.

s114
00:11:27.920 --> 00:11:29.839
You can connect to Z code.

s115
00:11:29.839 --> 00:11:34.880
Actually, the I think the operation is is similar to Kodak.

s116
00:11:34.880 --> 00:11:46.920
Actually, you you can try some techniques like go or other like compact technique techniques like what you've done in in Kodak and car code.

s117
00:11:46.920 --> 00:11:51.720
And this harness is I think it's the perfect one for GLM 102.

s118
00:11:51.720 --> 00:11:57.440
If you haven't experienced it, you can just search Z code or you can go to our booth.

s119
00:11:57.440 --> 00:12:01.880
We have our team members showing this harness to you.

s120
00:12:01.880 --> 00:12:04.680
And welcome to to our booth.

s121
00:12:04.680 --> 00:12:07.280
And welcome to like talk to me in the future.

s122
00:12:07.280 --> 00:12:11.840
And next time I'll definitely be in SF talking to everyone.

s123
00:12:11.840 --> 00:12:13.280
Yeah, thanks.

s124
00:12:13.280 --> 00:12:15.800
Thank you very much.

s125
00:12:17.480 --> 00:12:29.400
Um So uh trust me, the very first World's Fair, I wasn't allowed to come back in the country for my own conference, so I know exactly this feeling.

s126
00:12:29.400 --> 00:12:31.240
Uh but it it it's all good.

s127
00:12:31.240 --> 00:12:34.960
Uh the ZAI team, I really appreciate them uh making the effort.

s128
00:12:34.960 --> 00:12:35.720
They really want to meet you.

s129
00:12:35.720 --> 00:12:36.760
They're here to meet you.

s130
00:12:36.760 --> 00:12:45.680
And this is the whole point of the World's Fair, to bring all the top world's AI uh companies and labs uh all in one place, so you can do business together,

s131
00:12:45.680 --> 00:12:52.720
uh meet the people behind the models that you use, ask the questions that he cannot answer in public, but you can ask in private.

s132
00:12:52.720 --> 00:12:54.640
Um I'm also very proud.

s133
00:12:54.640 --> 00:13:00.080
Uh he showed uh Zhicheng showed that that list of uh Hugging Face uh you know, top uh contributors.

s134
00:13:00.080 --> 00:13:04.720
I think we are four for eight uh in that list uh present at World's Fair.

s135
00:13:04.720 --> 00:13:16.160
Uh we're working with Nvidia and Unsloth and Ollama and then uh all those other fine-tuners as well to uh to make our our sort of the inference and the local tracks uh that you're going to see over the next few days.

s136
00:13:16.160 --> 00:13:18.480
Um with that, thank you so much to Zhicheng.

s137
00:13:18.480 --> 00:13:21.520
Uh I'm going to invite on uh the next uh couple folks.

s138
00:13:21.520 --> 00:13:28.200
I think I I I think I might be doing Ali's job here, so uh we'll we'll talk to Hugging Face next and Minimax.

s139
00:13:28.200 --> 00:13:30.440
Thank you.